Preserve word boxes before tuning merged CDL OCR lines

When line OCR merges AAMVA field designators with values, preserve word-level geometry before tuning models

For photographed driver licenses, line-level OCR boxes can legitimately combine field designators, values, and neighboring fields. Enable PaddleOCR word boxes and retain them through normalization before adding substring repairs. First test a card crop or perspective rectification, then a small detector grid for scale, box threshold, and unclip ratio. Word boxes improve association but may still group contiguous ASCII; structured CDL output therefore still needs an AAMVA-aware parser and independently validated field rules. Do not treat detector tuning as semantic extraction.

Applies when: extracting structured fields from photographed AAMVA driver licenses with merged OCR lines. Keywords: PaddleOCR AAMVA CDL word boxes geometry unclip crop parser